diff --git a/spacy/_ml.pxd b/spacy/_ml.pxd
index b670b0f11..95afac99c 100644
--- a/spacy/_ml.pxd
+++ b/spacy/_ml.pxd
@@ -12,25 +12,34 @@ from .typedefs cimport hash_t, id_t
 from .tokens cimport Tokens
 
 
+cdef int arg_max(const weight_t* scores, const int n_classes) nogil
+
+
 cdef class Model:
-    cdef weight_t* score(self, atom_t* context) except NULL
-    cdef class_t predict(self, atom_t* context) except *
-    cdef class_t predict_among(self, atom_t* context, bint* valid) except *
-    cdef class_t predict_and_update(self, atom_t* context, const bint* valid,
-                                    const int* costs) except *
- 
+    cdef Pool mem
+    cdef int n_classes
+
+    cdef int update(self, atom_t* context, class_t guess, class_t gold, int cost) except -1
+
     cdef object model_loc
     cdef Extractor _extractor
     cdef LinearModel _model
 
+    cdef inline const weight_t* score(self, atom_t* context):
+        cdef int n_feats
+        feats = self._extractor.get_feats(context, &n_feats)
+        return self._model.get_scores(feats, n_feats)
+
 
 cdef class HastyModel:
-    cdef class_t predict(self, atom_t* context) except *
-    cdef class_t predict_among(self, atom_t* context, bint* valid) except *
-    cdef class_t predict_and_update(self, atom_t* context, const bint* valid,
-                                    const int* costs) except *
+    cdef Pool mem
+    cdef weight_t* _scores
  
-    cdef weight_t confidence
+    cdef const weight_t* score(self, atom_t* context) except NULL
+    cdef int update(self, atom_t* context, class_t guess, class_t gold, int cost) except -1
+
     cdef int n_classes
     cdef Model _hasty
     cdef Model _full
+    cdef readonly int hasty_cnt
+    cdef readonly int full_cnt
diff --git a/spacy/_ml.pyx b/spacy/_ml.pyx
index 9f26af003..7ad4bee37 100644
--- a/spacy/_ml.pyx
+++ b/spacy/_ml.pyx
@@ -4,7 +4,6 @@ from __future__ import division
 
 from os import path
 import os
-from collections import defaultdict
 import shutil
 import random
 import json
@@ -13,80 +12,39 @@ import cython
 from thinc.features cimport Feature, count_feats
 
 
-def setup_model_dir(tag_names, tag_map, templates, model_dir):
-    if path.exists(model_dir):
-        shutil.rmtree(model_dir)
-    os.mkdir(model_dir)
-    config = {
-        'templates': templates,
-        'tag_names': tag_names,
-        'tag_map': tag_map
-    }
-    with open(path.join(model_dir, 'config.json'), 'w') as file_:
-        json.dump(config, file_)
+cdef int arg_max(const weight_t* scores, const int n_classes) nogil:
+    cdef int i
+    cdef int best = 0
+    cdef weight_t mode = scores[0]
+    for i in range(1, n_classes):
+        if scores[i] > mode:
+            mode = scores[i]
+            best = i
+    return best
 
 
 cdef class Model:
     def __init__(self, n_classes, templates, model_loc=None):
+        if model_loc is not None and path.isdir(model_loc):
+            model_loc = path.join(model_loc, 'model')
+        self.mem = Pool()
+        self.n_classes = n_classes
         self._extractor = Extractor(templates)
         self._model = LinearModel(n_classes, self._extractor.n_templ)
         self.model_loc = model_loc
         if self.model_loc and path.exists(self.model_loc):
             self._model.load(self.model_loc, freq_thresh=0)
 
-    cdef const weight_t* score(self, atom_t* context) except NULL:
+    cdef int update(self, atom_t* context, class_t guess, class_t gold, int cost) except -1:
         cdef int n_feats
-        cdef const Feature* feats = self._extractor.get_feats(context, &n_feats)
-        return self._model.get_scores(feats, n_feats)
-
-    cdef class_t predict(self, atom_t* context) except *:
-        cdef weight_t _
-        scores = self.score(context)
-        guess = _arg_max(scores, self._model.nr_class, &_)
-        return guess
-
-    cdef class_t predict_among(self, atom_t* context, const bint* valid) except *:
-        cdef weight_t _
-        scores = self.score(context)
-        return _arg_max_among(scores, valid, self._model.nr_class, &_)
-
-    cdef class_t predict_and_update(self, atom_t* context, const bint* valid,
-                                    const int* costs) except *:
-        cdef:
-            int n_feats
-            const Feature* feats
-            const weight_t* scores
-
-            int guess
-            int best
-            int cost
-            int i
-            weight_t score
-            weight_t _
-        
-        feats = self._extractor.get_feats(context, &n_feats)
-        scores = self._model.get_scores(feats, n_feats)
-        guess = _arg_max_among(scores, valid, self._model.nr_class, &_)
-        cost = costs[guess]
         if cost == 0:
             self._model.update({})
-            return guess
-
-        guess_counts = defaultdict(int)
-        best_counts = defaultdict(int)
-        for i in range(n_feats):
-            feat = (feats[i].i, feats[i].key)
-            upd = feats[i].value * cost
-            best_counts[feat] += upd
-            guess_counts[feat] -= upd
-        best = -1
-        score = 0
-        for i in range(self._model.nr_class):
-            if valid[i] and costs[i] == 0 and (best == -1 or scores[i] > score):
-                best = i
-                score = scores[i]
-        self._model.update({guess: guess_counts, best: best_counts})
-        return guess
+        else:
+            feats = self._extractor.get_feats(context, &n_feats)
+            counts = {gold: {}, guess: {}}
+            count_feats(counts[gold], feats, n_feats, cost)
+            count_feats(counts[guess], feats, n_feats, -cost)
+            self._model.update(counts)
 
     def end_training(self):
         self._model.end_training()
@@ -94,41 +52,34 @@ cdef class Model:
 
 
 cdef class HastyModel:
-    def __init__(self, n_classes, hasty_templates, full_templates, model_dir,
-                 weight_t confidence=0.1):
+    def __init__(self, n_classes, hasty_templates, full_templates, model_dir):
+        full_templates = tuple([t for t in full_templates if t not in hasty_templates])
+        self.mem = Pool()
         self.n_classes = n_classes
-        self.confidence = confidence
+        self._scores = <weight_t*>self.mem.alloc(self.n_classes, sizeof(weight_t))
+        assert path.exists(model_dir)
+        assert path.isdir(model_dir)
         self._hasty = Model(n_classes, hasty_templates, path.join(model_dir, 'hasty_model'))
         self._full = Model(n_classes, full_templates, path.join(model_dir, 'full_model'))
+        self.hasty_cnt = 0
+        self.full_cnt = 0
 
-    cdef class_t predict(self, atom_t* context) except *:
-        cdef weight_t ratio
-        scores = self._hasty.score(context)
-        guess = _arg_max(scores, self.n_classes, &ratio)
-        if ratio < self.confidence:
-            return guess
+    cdef const weight_t* score(self, atom_t* context) except NULL:
+        cdef int i
+        hasty_scores = self._hasty.score(context)
+        if will_use_hasty(hasty_scores, self._hasty.n_classes):
+            self.hasty_cnt += 1
+            return hasty_scores
         else:
-            return self._full.predict(context)
+            self.full_cnt += 1
+            full_scores = self._full.score(context)
+            for i in range(self.n_classes):
+                self._scores[i] = full_scores[i] + hasty_scores[i]
+            return self._scores
 
-    cdef class_t predict_among(self, atom_t* context, bint* valid) except *:
-        cdef weight_t ratio
-        scores = self._hasty.score(context)
-        guess = _arg_max_among(scores, valid, self.n_classes, &ratio)
-        if ratio < self.confidence:
-            return guess
-        else:
-            return self._full.predict(context)
-
-    cdef class_t predict_and_update(self, atom_t* context, bint* valid, int* costs) except *:
-        cdef weight_t ratio
-        scores = self._hasty.score(context)
-        _arg_max_among(scores, valid, self.n_classes, &ratio)
-        hasty_guess = self._hasty.predict_and_update(context, valid, costs)
-        full_guess = self._full.predict_and_update(context, valid, costs)
-        if ratio < self.confidence:
-            return hasty_guess
-        else:
-            return full_guess
+    cdef int update(self, atom_t* context, class_t guess, class_t gold, int cost) except -1:
+        self._hasty.update(context, guess, gold, cost)
+        self._full.update(context, guess, gold, cost)
 
     def end_training(self):
         self._hasty.end_training()
@@ -136,31 +87,29 @@ cdef class HastyModel:
 
 
 @cython.cdivision(True)
-cdef int _arg_max(const weight_t* scores, int n_classes, weight_t* ratio) except -1:
-    cdef int best = 0
-    cdef weight_t score = scores[best]
+cdef bint will_use_hasty(const weight_t* scores, int n_classes) nogil:
+    cdef:
+        weight_t best_score, second_score
+        int best, second
+
+    if scores[0] >= scores[1]:
+        best = 0
+        best_score = scores[0]
+        second = 1
+        second_score = scores[1]
+    else:
+        best = 1
+        best_score = scores[1]
+        second = 0
+        second_score = scores[0]
     cdef int i
-    ratio[0] = 0.0
-    for i in range(1, n_classes):
-        if scores[i] >= score:
-            if score > 0:
-                ratio[0] = score / scores[i]
-            score = scores[i]
+    for i in range(2, n_classes):
+        if scores[i] > best_score:
+            second_score = best_score
+            second = best
             best = i
-    return best
-
-
-@cython.cdivision(True)
-cdef int _arg_max_among(const weight_t* scores, const bint* valid, int n_classes,
-                        weight_t* ratio) except -1:
-    cdef int clas
-    cdef weight_t score = 0
-    cdef int best = -1
-    ratio[0] = 0
-    for clas in range(n_classes):
-        if valid[clas] and (best == -1 or scores[clas] > score):
-            if score > 0:
-                ratio[0] = score / scores[clas]
-            score = scores[clas]
-            best = clas
-    return best
+            best_score = scores[i]
+        elif scores[i] > second_score:
+            second_score = scores[i]
+            second = i
+    return best_score > 0 and second_score < (best_score / 2)
diff --git a/spacy/en/__init__.py b/spacy/en/__init__.py
index 80efb9cad..df2b26b42 100644
--- a/spacy/en/__init__.py
+++ b/spacy/en/__init__.py
@@ -82,16 +82,13 @@ class English(object):
             tokens (spacy.tokens.Tokens):
         """
         tokens = self.tokenizer.tokenize(text)
-        if self.tagger and tag:
+        if tag:
             self.tagger(tokens)
-        if self.parser and parse:
+        if parse:
             self.parser.parse(tokens)
         return tokens
 
     @property
     def tags(self):
         """List of part-of-speech tag names."""
-        if self.tagger is None:
-            return []
-        else:
-            return self.tagger.tag_names
+        return self.tagger.tag_names
diff --git a/spacy/en/pos.pyx b/spacy/en/pos.pyx
index f8ebc146a..67719ceb0 100644
--- a/spacy/en/pos.pyx
+++ b/spacy/en/pos.pyx
@@ -1,11 +1,13 @@
 # cython: profile=True
 from os import path
 import json
+import os
+import shutil
 
 from libc.string cimport memset
 
 from cymem.cymem cimport Address
-from thinc.typedefs cimport atom_t
+from thinc.typedefs cimport atom_t, weight_t
 
 from ..typedefs cimport univ_tag_t
 from ..typedefs cimport NO_TAG, ADJ, ADV, ADP, CONJ, DET, NOUN, NUM, PRON, PRT, VERB
@@ -14,6 +16,8 @@ from ..typedefs cimport id_t
 from ..structs cimport TokenC, Morphology, Lexeme
 from ..tokens cimport Tokens
 from ..morphology cimport set_morph_from_dict
+from .._ml cimport arg_max
+
 from .lemmatizer import Lemmatizer
 
 
@@ -206,6 +210,19 @@ cdef struct _CachedMorph:
     int lemma
 
 
+def setup_model_dir(tag_names, tag_map, templates, model_dir):
+    if path.exists(model_dir):
+        shutil.rmtree(model_dir)
+    os.mkdir(model_dir)
+    config = {
+        'templates': templates,
+        'tag_names': tag_names,
+        'tag_map': tag_map
+    }
+    with open(path.join(model_dir, 'config.json'), 'w') as file_:
+        json.dump(config, file_)
+
+
 cdef class EnPosTagger:
     """A part-of-speech tagger for English"""
     def __init__(self, StringStore strings, data_dir):
@@ -218,8 +235,8 @@ cdef class EnPosTagger:
         self.tag_map = cfg['tag_map']
         cdef int n_tags = len(self.tag_names) + 1
 
-        self.model = Model(n_tags, cfg['templates'], model_dir=model_dir)
-
+        hasty_templates = ((W_sic,), (P1_pos, P2_pos), (N1_sic,))
+        self.model = Model(n_tags, cfg['templates'], model_dir)
         self._morph_cache = PreshMapArray(n_tags)
         self.tags = <PosTag*>self.mem.alloc(n_tags, sizeof(PosTag))
         for i, tag in enumerate(sorted(self.tag_names)):
@@ -239,30 +256,27 @@ cdef class EnPosTagger:
         """
         cdef int i
         cdef atom_t[N_CONTEXT_FIELDS] context
-        cdef TokenC* t = tokens.data
+        cdef const weight_t* scores
         for i in range(tokens.length):
-            if t[i].fine_pos == 0:
-                fill_context(context, i, t)
-                t[i].fine_pos = self.model.predict(context)
-                self.set_morph(i, t)
+            if tokens.data[i].fine_pos == 0:
+                fill_context(context, i, tokens.data)
+                scores = self.model.score(context)
+                tokens.data[i].fine_pos = arg_max(scores, self.model.n_classes)
+                self.set_morph(i, tokens.data)
 
-    def train(self, Tokens tokens, py_golds):
+    def train(self, Tokens tokens, object golds):
         cdef int i
         cdef atom_t[N_CONTEXT_FIELDS] context
-        cdef Address costs_mem = Address(self.n_tags, sizeof(int))
-        cdef Address valid_mem = Address(self.n_tags, sizeof(bint))
-        cdef int* costs = <int*>costs_mem.ptr
-        cdef bint* valid = <bint*>valid_mem.ptr
-        memset(valid, 1, sizeof(int) * self.n_tags)
+        cdef const weight_t* scores
         correct = 0
-        cdef TokenC* t = tokens.data
         for i in range(tokens.length):
-            fill_context(context, i, t)
-            memset(costs, 1, sizeof(int) * self.n_tags)
-            costs[py_golds[i]] = 0
-            t[i].fine_pos = self.model.predict_and_update(context, valid, costs)
-            self.set_morph(i, t)
-            correct += costs[t[i].fine_pos] == 0
+            fill_context(context, i, tokens.data)
+            scores = self.model.score(context)
+            guess = arg_max(scores, self.model.n_classes)
+            self.model.update(context, guess, golds[i], guess != golds[i])
+            tokens.data[i].fine_pos = guess
+            self.set_morph(i, tokens.data)
+            correct += guess == golds[i]
         return correct
 
     cdef int set_morph(self, const int i, TokenC* tokens) except -1:
diff --git a/spacy/syntax/_parse_features.pyx b/spacy/syntax/_parse_features.pyx
index 5c7f39ff7..bfdc94029 100644
--- a/spacy/syntax/_parse_features.pyx
+++ b/spacy/syntax/_parse_features.pyx
@@ -85,7 +85,6 @@ cdef int fill_context(atom_t* context, State* state) except -1:
             if state.stack_len >= 3:
                 context[S2_has_head] = has_head(get_s2(state))
 
-
 unigrams = (
     (S2W, S2p),
     (S2c6, S2p),
@@ -347,6 +346,9 @@ clusters = (
 )
 
 
+hasty = s0_n0 + n0_n1 + trigrams
+
+
 def pos_bigrams():
     kernels = [S2w, S1w, S0w, S0lw, S0rw, N0w, N0lw, N1w]
     bitags = []
diff --git a/spacy/syntax/parser.pxd b/spacy/syntax/parser.pxd
index 224ea490b..e4fc38969 100644
--- a/spacy/syntax/parser.pxd
+++ b/spacy/syntax/parser.pxd
@@ -1,5 +1,4 @@
-from thinc.features cimport Extractor
-from thinc.learner cimport LinearModel
+from .._ml cimport Model, HastyModel
 
 from .arc_eager cimport TransitionSystem
 
@@ -8,8 +7,7 @@ from ..tokens cimport Tokens, TokenC
 
 cdef class GreedyParser:
     cdef object cfg
-    cdef Extractor extractor
-    cdef readonly LinearModel model
+    cdef readonly Model model
     cdef TransitionSystem moves
 
     cpdef int parse(self, Tokens tokens) except -1
diff --git a/spacy/syntax/parser.pyx b/spacy/syntax/parser.pyx
index 9b23cea1c..755766eb3 100644
--- a/spacy/syntax/parser.pyx
+++ b/spacy/syntax/parser.pyx
@@ -7,7 +7,7 @@ cimport cython
 from libc.stdint cimport uint32_t, uint64_t
 import random
 import os.path
-from os.path import join as pjoin
+from os import path
 import shutil
 import json
 
@@ -52,26 +52,23 @@ cdef unicode print_state(State* s, list words):
 def get_templates(name):
     pf = _parse_features
     if name == 'zhang':
-        return pf.arc_eager
+        return pf.unigrams, pf.arc_eager
     else:
-        return pf.unigrams + pf.s0_n0 + pf.s1_n0 + pf.s0_n1 + pf.n0_n1 + \
-               pf.tree_shape + pf.trigrams
+        return pf.hasty, (pf.unigrams + pf.s0_n0 + pf.s1_n0 + pf.s0_n1 + pf.n0_n1 + \
+                             pf.tree_shape + pf.trigrams)
 
 
 cdef class GreedyParser:
     def __init__(self, model_dir):
         assert os.path.exists(model_dir) and os.path.isdir(model_dir)
         self.cfg = Config.read(model_dir, 'config')
-        self.extractor = Extractor(get_templates(self.cfg.features))
         self.moves = TransitionSystem(self.cfg.left_labels, self.cfg.right_labels)
-        self.model = LinearModel(self.moves.n_moves, self.extractor.n_templ)
-        if os.path.exists(pjoin(model_dir, 'model')):
-            self.model.load(pjoin(model_dir, 'model'))
+        hasty_templ, full_templ = get_templates(self.cfg.features)
+        #self.model = HastyModel(self.moves.n_moves, hasty_templ, full_templ, model_dir)
+        self.model = Model(self.moves.n_moves, full_templ, model_dir)
 
     cpdef int parse(self, Tokens tokens) except -1:
         cdef:
-            const Feature* feats
-            const weight_t* scores
             Transition guess
             uint64_t state_key
 
@@ -81,8 +78,7 @@ cdef class GreedyParser:
         cdef State* state = init_state(mem, tokens.data, tokens.length)
         while not is_final(state):
             fill_context(context, state)
-            feats = self.extractor.get_feats(context, &n_feats)
-            scores = self.model.get_scores(feats, n_feats)
+            scores = self.model.score(context)
             guess = self.moves.best_valid(scores, state)
             self.moves.transition(state, &guess)
         return 0
@@ -106,35 +102,13 @@ cdef class GreedyParser:
         
         cdef State* state = init_state(mem, tokens.data, tokens.length)
         while not is_final(state):
-            fill_context(context, state) 
-            feats = self.extractor.get_feats(context, &n_feats)
-            scores = self.model.get_scores(feats, n_feats)
+            fill_context(context, state)
+            scores = self.model.score(context)
             guess = self.moves.best_valid(scores, state)
             best = self.moves.best_gold(&guess, scores, state, heads_array, labels_array)
-            counts = _get_counts(guess.clas, best.clas, feats, n_feats, guess.cost)
-            self.model.update(counts)
+            self.model.update(context, guess.clas, best.clas, guess.cost)
             self.moves.transition(state, &guess)
         cdef int n_corr = 0
         for i in range(tokens.length):
             n_corr += (i + state.sent[i].head) == gold_heads[i]
         return n_corr
-
-
-cdef dict _get_counts(int guess, int best, const Feature* feats, const int n_feats,
-                      int inc):
-    if guess == best:
-        return {}
-
-    gold_counts = {}
-    guess_counts = {}
-    cdef int i
-    for i in range(n_feats):
-        key = (feats[i].i, feats[i].key)
-        if key in gold_counts:
-            gold_counts[key] += (feats[i].value * inc)
-            guess_counts[key] -= (feats[i].value * inc)
-        else:
-            gold_counts[key] = (feats[i].value * inc)
-            guess_counts[key] = -(feats[i].value * inc)
-    return {guess: guess_counts, best: gold_counts}
-